On Compensation Primitives as Adaptable Processes

Transcription

1 On Compensation Primitives as Adaptable Processes Jovana Dedeić University of Novi Sad Jovanka Pantović (Novi Sad) and Jorge A. Pérez (Groningen) LAP Dubrovnik, September 24, 2015

2 Outline 1 Context Introduction Compensable Processes Adaptable Processes 2 The Encoding Basic Intuition Formal definition of the encoding The Encoding: By Example

3 Introduction Many distributed software applications exploit long-running transactions (LRTs). One particularly delicate aspect of LRTs management is handling (partial) failures

4 Introduction Many distributed software applications exploit long-running transactions (LRTs). One particularly delicate aspect of LRTs management is handling (partial) failures The last decade has seen the emergence of specialized constructs, such as exceptions and compensations. In this work, we study process calculi with constructs for compensations.

5 Compensable Processes (CPs) Several calculi with compensations have been proposed. The calculus of Lanese et al (ESOP 10) extends the π-calculus with: t[p, Q] Q inst λx.r.p transaction scopes protected blocks compensation updates Also, output prefixes t represent abortion signals.

6 Compensable Processes (CPs) Several calculi with compensations have been proposed. The calculus of Lanese et al (ESOP 10) extends the π-calculus with: t[p, Q] Q inst λx.r.p transaction scopes protected blocks compensation updates Also, output prefixes t represent abortion signals. Several Labeled Transition Systems (LTS) define semantics. Example: given process P = t[t 1 [P 1, Q 1 ] R P 2, Q 2 ] t In the discarding semantics, we have P τ D P 2 Q 2.

7 Adaptable Processes (APs) A process calculi approach to evolvability, in a broad sense. Proposed by Bravetti et al (FORTE 11, LMCS 12). Studied from several perspectives, e.g., expressiveness, decidability/verification, session types (SAC 13, WSFM 14). Runtime modifications to (located) process behaviors, upon exceptional circumstances not necessarily negative. Simple formulation: higher-order process passing.

8 Adaptable Processes (APs) The calculus extends CCS with locations l, l,... and l{(x).q} l[p ] update prefix located process

9 Adaptable Processes (APs) The calculus extends CCS with locations l, l,... and l{(x).q} l[p ] update prefix located process These two constructs are meant to synchronize. Located processes are transparent and can be arbitrarily nested. This is useful to structure processes into hierarchies.

10 Adaptable Processes (APs) The calculus extends CCS with locations l, l,... and l{(x).q} l[p ] update prefix located process These two constructs are meant to synchronize. Located processes are transparent and can be arbitrarily nested. This is useful to structure processes into hierarchies. Simple reduction semantics (C,D and E are evaluation contexts): [ E C [ a.p ] D [ a.q ]] E [C [ P ] D [ Q ]] [ E C [ l[p ] ] D [ l{(x).q}.r ]] E [C [ Q{P/X} ] D [ R ]]

11 CPs and APs: Similarities and Differences A transaction scope reacts to an abortion signal (an output) by removing the default behavior and running its compensation. Assuming no protected blocks in P we have: t t[p, Q] τ D Q Similarly, a process located at l reacts to a synchronization with an update prefix for l. Assuming X fv(q), we have: l[p ] l{(x).q} Q

12 CPs and APs: Similarities and Differences A transaction scope reacts to an abortion signal (an output) by removing the default behavior and running its compensation. Assuming no protected blocks in P we have: t t[p, Q] τ D Q Similarly, a process located at l reacts to a synchronization with an update prefix for l. Assuming X fv(q), we have: l[p ] l{(x).q} Q Some differences: 1. In CPs, a transaction scope couples a default behavior and its associated compensation. In APs, update prefixes and located processes are defined separately. 2. In APs, there is no notion of protected block.

13 Our Contribution: Encoding CPs into APs We have encoded CPs (with different semantics for failure) into APs. Our encodings not only are a non trivial application of process mobility. They shed light on the intricate semantics of compensable processes. The main challenge to encodability is in representing the different failure semantics using adaptable process.

14 Our Contribution: Motivation Our motivation is twofold. First Understanding how different semantics for compensable processes can be uniformly implemented as adaptable processes. Second Our encodings could enable the transference of, e.g., decidability results or type systems from adaptable processes to calculi with compensations.

15 Compensable Processes end Adaptable Processes Both CPs and APs are defined as variants of CCS. The syntax of the calculus of compensable processes. π ::= a a P, Q ::= 0 π.p!p (νa)p P Q t[p, Q] Q X inst λx.r.p The syntax of the calculus of adaptable processes. π ::= a a l{(x).q} P ::= 0 π.p!p P Q (νa)p l[p ] X

16 The Encoding: Basic Intuition We roughly encode protected blocks and transactions as: R t,ρ = p t,ρ [ R ɛ ] t[p, Q] ρ = t [ P t,ρ ] }{{} (a) l t.π 1..π k.p t [ Q t,ρ ] }{{} (b) t.l t.k }{{} (c) Paths ρ describe the structure of nested transactions Protected blocks are placed in designated locations p t. Part (a) is a located process encoding the default activity Part (b) represents the compensation activity and is protected by special prefixes (π 1..π k ). Part (c) handles abortion signals, collecting protected blocks. It is meant to consume prefixes π 1,, π k.

17 The Encoding: Discarding Semantics Let P be a compensable process and let ρ be a path. The encoding D ρ of compensable processes into adaptable processes is defined as follows: [ ] D P ρ = p ρ D P ɛ ] D t[p, Q] ρ = t [D P t,ρ D 0 ρ = 0 D P 1 P 2 ρ = D P 1 ρ D P 2 ρ D π.p ρ = π.d P ρ D! P ρ =! D P ρ D (νa)p ρ = (νa)d P ρ D Q npb D (P ) t,ρ t.l t.k t.0

18 The Encoding: By Example Let P 0 = t[r P, Q] t be a CP with npb D (R) = 0. τ Then P 0 D P Q. We obtain: ] D P 0 ɛ = t [D R P t,ɛ D Q 1 t,ɛ t.l t.k t t

19 The Encoding: By Example Let P 0 = t[r P, Q] t be a CP with npb D (R) = 0. τ Then P 0 D P Q. We obtain: ] D P 0 ɛ = t [D R P t,ɛ D Q 1 t,ɛ t.l t.k t t [ ] = t [D R ] t,ɛ p t,ɛ D P ɛ { l t.p t,ɛ (X).z { [ ]} } p ɛ [X] m t.p ɛ D Q ɛ.(z[0] m t.k t.t{ }) t.l t.k t t

20 The Encoding: By Example Let P 0 = t[r P, Q] t be a CP with npb D (R) = 0. τ Then P 0 D P Q. We obtain: ] D P 0 ɛ = t [D R P t,ɛ D Q 1 t,ɛ t.l t.k t t [ ] = t [D R ] t,ɛ p t,ɛ D P ɛ { l t.p t,ɛ (X).z { [ ]} } p ɛ [X] m t.p ɛ D Q ɛ.(z[0] m t.k t.t{ }) t.l t.k t t t [D R t,ɛ z { [ ]} ] p ɛ [D P ɛ ] m t.p ɛ D Q ɛ z[0] m t.k t.t { (Y ).0 } k t

21 The Encoding: By Example Let P 0 = t[r P, Q] t be a CP with npb D (R) = 0. τ Then P 0 D P Q. We obtain: ] D P 0 ɛ = t [D R P t,ɛ D Q 1 t,ɛ t.l t.k t t [ ] = t [D R ] t,ɛ p t,ɛ D P ɛ { l t.p t,ɛ (X).z { [ ]} } p ɛ [X] m t.p ɛ D Q ɛ.(z[0] m t.k t.t{ }) t.l t.k t t t [D R t,ɛ z { [ ]} ] p ɛ [D P ɛ ] m t.p ɛ D Q ɛ z[0] m t.k t.t { (Y ).0 } [ ] [ ] k t p ɛ D P ɛ pɛ D Q ɛ = D P Q ɛ

22 The Encoding: Operational Correspondence Theorem Let P be a compensable process and let ρ be a path. a) If P τ D P then D P ρ D P ρ b) If D P ρ Q then P s.t. P τ D P and Q D P ρ.

23 Further Results in the paper We have described CPs with static recovery (no compensation updates inst λx.r.p ) and discarding semantics.

24 Further Results in the paper We have described CPs with static recovery (no compensation updates inst λx.r.p ) and discarding semantics. In the paper we also consider encodings for CPs with two further failure semantics: preserving and aborting.

25 Further Results in the paper We have described CPs with static recovery (no compensation updates inst λx.r.p ) and discarding semantics. In the paper we also consider encodings for CPs with two further failure semantics: preserving and aborting. Example: let P be a CP t[t 1 [P 1, Q 1 ] R P 2, Q 2 ] t. In the preserving and aborting semantics, we have the internal transitions: P P τ P t 1 [P 1, Q 1 ] P 2 Q 2 τ A Q 1 P 2 Q 2 The encodings into APs and proofs follow similar principles.

26 Further Results in the paper We have described CPs with static recovery (no compensation updates inst λx.r.p ) and discarding semantics. In the paper we also consider encodings for CPs with two further failure semantics: preserving and aborting. Example: let P be a CP t[t 1 [P 1, Q 1 ] R P 2, Q 2 ] t. In the preserving and aborting semantics, we have the internal transitions: P P τ P t 1 [P 1, Q 1 ] P 2 Q 2 τ A Q 1 P 2 Q 2 The encodings into APs and proofs follow similar principles. We also cover dynamic recovery, which includes compensation updates inst λx.r.p.

27 The paper Jovana Dedeic, Jovanka Pantovic, Jorge A. Pérez: On Compensation Primitives as Adaptable Processes. EXPRESS/SOS 2015: 16-30

28 Future Plans We plan to consider the reverse direction of encoding. Cast our encodability results into a setting with session types: The source language could be the typed calculus with interactional exceptions (Carbone et al, CONCUR 08 ) The target language could be recently proposed extensions of adaptable processes with session types (SAC 13, WSFM 14).

29 On Compensation Primitives as Adaptable Processes Jovana Dedeić University of Novi Sad Jovanka Pantović (Novi Sad) and Jorge A. Pérez (Groningen) LAP Dubrovnik, September 24, 2015